Reasoning-based AI medical assistant offers a glimpse of the future of pediatric care

Google Cloud Results

Provides full-context insight into patient health histories to improve speed, accuracy, and compassion in pediatric care
Reduces model training time without increasing costs
Eliminates reliance on retrieval-based systems that take precious time and may fail to interpret full context

Children's Hospital of Philadelphia used secure Google Cloud Trillium TPUs to train a reasoning AI assistant that aims to give researchers and, one day, physicians deep insight from patient histories to help accelerate and improve pediatric care.

When the stakes are as high as a child's health, accuracy and speed in medical care are non-negotiable — but time pressures and fragmented patient data create a barrier to providing optimal pediatric care. While technologies exist to help physicians comb through dense repositories of electronic health records (EHRs), most systems fail to understand the full context of a patient's health history, which can slow down essential diagnosis and treatment processes.

Researchers at the Children's Hospital of Philadelphia (CHOP) imagined a future where technology could overcome these obstacles and help physicians focus on providing precise, compassionate, and timely care.

"We asked ourselves, what if there was an AI medical assistant that could help physicians better understand their patients? An AI that could answer any question about a patient that's sitting in front of them?," says Dr. Ian Campbell, a pediatric medical geneticist and assistant professor of pediatrics at CHOP. "That's what we set out to build."

With Google Cloud , Dr. Campbell and his team created a secure, compliant AI assistant that has the potential to one day give physicians deep, contextual insight into patient health histories to improve the speed and quality of pediatric medicine.

We asked ourselves, what if there was an AI medical assistant that could help physicians better understand their patients? An AI that could answer any question about a patient that's sitting in front of them? That's what we set out to build.

Dr. Ian Campbell

Assistant Professor of Pediatrics, CHOP

Navigating clinical data overload while protecting patient privacy

In one instance, a RAG system incorrectly reported that a patient had never been prescribed antiarrhythmic medications, despite explicit references to such medications in the EHR. To overcome these limitations and inaccuracies, we decided to pre-train our model to learn about the patient in advance while protecting their privacy.

Dr. Ian Campbell

Assistant Professor of Pediatrics, CHOP

Physicians rely on information from EHRs to deliver appropriate patient care, but manually reviewing the entirety of each patient's health history isn't possible in the always-on world of a pediatric hospital. Many medical facilities are experimenting with AI-based systems that rely on retrieval-augmented generation (RAG) to pull relevant information from EHRs, but these solutions are costly and slow. They also struggle to interpret medical context within EHRs, which can contain millions of data points about individual patients. Moreover, the lack of model learning limits RAG systems to brief, incomplete impressions during inference.

"In one instance, a RAG system incorrectly reported that a patient had never been prescribed antiarrhythmic medications, despite explicit references to such medications in the EHR," Dr. Campbell explains. "To overcome these limitations and inaccuracies, we decided to pre-train our model to learn about the patient in advance while protecting their privacy."

However, as a nonprofit hospital, CHOP lacked the necessary AI accelerators and on-premises computational capacity to train such an ambitious model. Dealing with private health information also required strict compliance with Health Insurance Portability and Accountability Act (HIPAA) regulations. To address these infrastructure gaps while safeguarding patient data, Dr. Campbell and his team turned to Google Cloud.

Building a patient-specific AI assistant with MaxText on Google Cloud

Instead of relying on retrieval-based systems, Dr. Campbell's team developed a reasoning-based model capable of patient-specific understanding. This AI assistant uses models based on Llama 3.3 70B and other advanced architectures to effectively pre-learn the medical histories of over 1.6 million pediatric patients, using 146 million clinical notes as training data. The team used open-source JAX and MaxText to train their AI assistant with Trillium TPUs from Google Cloud, partially supported by the TPU Research Cloud .

"MaxText includes all the essential features of LLM training, including resumable deterministic data loading, fully customizable gradient re-materialization, CPU offload, and asynchronized emergency checkpointing," Dr. Campbell notes. "It made transitioning to the cloud very straightforward."

Privacy and security are our top priorities at CHOP. Our institutional review board is there to protect our human subjects, while our AI governance committee ensures that we are using the technology responsibly.

Dr. Ian Campbell

Assistant Professor of Pediatrics, CHOP

Given the sensitivity of pediatric medical data, CHOP's institutional review board and AI governance committees supervised the project. All model training and inference occurred in the strictly controlled, HIPAA-compliant environment within Google Cloud to ensure patient privacy.

"Privacy and security are our top priorities at CHOP," Dr. Campbell says. "Our institutional review board is there to protect our human subjects, while our AI governance committee ensures that we are using the technology responsibly."

Within this secure and compliant infrastructure, the AI assistant retains deep contextual knowledge about each patient, enabling fast and accurate responses in clinically relevant settings. For example, it showed powerful reasoning capabilities when answering why a patient required a liver transplant. Using only the patient's name, date of birth, and EHR number, it recalled everything it knew about the patient, including rare genetic conditions, metabolic crises, liver biopsies, and subsequent diagnoses.

"This was the first question I asked our new reasoning model," Dr. Campbell recalls. "I immediately thought, 'This might be the future of medicine.'"

Improving pediatric healthcare through collaborative innovation

As a clinician, I'm excited about a future where AI can help me take better care of my patients. I look forward to bringing this innovation to the industry to help improve pediatric care everywhere.

Dr. Ian Campbell

Assistant Professor of Pediatrics, CHOP

Where traditional RAG systems sometimes failed in the past, CHOP's reasoning-based AI assistant succeeds for the future. Its ability to process and recall patient-specific histories will one day help clinicians make quick, informed decisions. By acting as a collaborative partner to physicians, the assistant provides a glimpse of a new standard for operational efficiency in patient care.

"In addition to understanding deep context from a patient's history, our AI assistant learns the culture and practices of the health system that it serves," Dr. Campbell explains. "It can learn how to implement our hospital's treatment pathways or the styles and preferences of individual clinicians."

The model currently supports CHOP's research projects by delivering insights on participants while always prioritizing patient privacy and aligning with compliance frameworks. And by leveraging the computational power of Trillium TPUs, the CHOP team reduced the time to train its model at a nonprofit-friendly price point.

While this AI assistant already provides promising results within CHOP's ecosystem, the team aims to collaborate with other institutions to help them train their own models using their unique datasets. Efforts are also underway to expand the model from 70 billion parameters to even larger architectures, introduce reinforcement learning to improve emerging tasks, and explore other model types for more adaptive capabilities.

"As a clinician, I'm excited about a future where AI can help me take better care of my patients," Dr. Campbell says. "I look forward to bringing this innovation to the industry to help improve pediatric care everywhere."

Founded in 1855, the Children's Hospital of Philadelphia is the nation's first freestanding pediatric hospital, recognized for providing world-class care and industry-leading research.